+ THREE INSIGHTS FOR THE WEEK |
1. Artificial intelligence adoption is happening fastest among people who are highly paid and experienced, according to a new poll of 4,000 workers in the U.S. and the U.K.
More than 60% of the best-paid workers are daily AI users, compared with 16% of the lowest-paid earners, according to the Financial Times, which conducted the poll.
“The rhetoric out there is that the tools are going to be democratizing,” MIT economist Daron Acemoglu, a Nobel laureate, told the publication.
“But the reality is that … you require a certain degree of education, abstract and quantitative skills, familiarity with computers and coding in order to be using the models,” he said. “AI is going to increase inequality between labor and capital. That is almost for sure.”
The poll also found evidence of a gender divide, with men more likely to use AI than women.
|
|
|
|
2. Is today’s tech-heavy stock market dangerously overconcentrated? Maybe not. That’s the message from MIT Sloan senior lecturer Mark Kritzman, whose research counters a Wall Street scare story.
Many financial advisers are warning that the S&P 500’s outsize exposure to a handful of giant tech companies puts index fund investors at risk. Kritzman, the CEO of Windham Capital Management, told The Wall Street Journal that his data says otherwise. Looking back 90 years, he’s found that investors who reduced their stock holdings during periods of rising concentration and added to them during declining concentration earned an annualized average 0.9 percentage points less than those who simply held on to them.
Why? Because large companies have far more diversified economic exposures — suppliers, markets, technology, geography — than their index weight suggests. Kritzman has found that owning just a handful of the biggest companies carries essentially the same risk profile as owning all remaining S&P 500 stocks combined.
“The largest stocks are just safer,” Kritzman said.
|
|
|
|
3. When MIT postdoctoral researcher Max Fornerod set out to commercialize portable DNA-detection technology, his problem wasn’t a lack of applications — it was the abundance of them. The technology had potential uses across hospitals, restaurants, animal health, and at-home diagnostics.
Where to start? Fornerod and his team turned to the MIT Entrepreneurship JetPack, an AI tool developed by the Martin Trust Center for MIT Entrepreneurship. “JetPack helped us when we were stuck, when I ran out of ideas, when I was tired,” Fornerod said.
Parents emerged as an early potential market, and thus was born Sensopore, a diagnostic device to help families test at home for everyday illnesses like strep throat. It sends information to an app for connecting with telehealth providers and submitting prescriptions.
From there, JetPack helped the team build customer personas and pressure-test assumptions, including one that surprised them: First-time parents of newborns turned out to be a poor customer fit. “They don’t want to test their kids themselves,” Fornerod said. More experienced parents in suburban or rural areas without easy in-person access to doctors emerged as a market where the company could get a foothold.
|
|
|
|
How generative AI ‘persuasion bombs’ users — and how to fight back
|
When professionals try to validate AI outputs, generative AI responds not with corrections or candor but with escalating persuasion.
|
|
|
|
Artificial intelligence makes mistakes. That’s why it’s important for users to take time to validate and question outputs, keeping a human in the loop.
But new research from MIT Sloan professor Kate Kellogg and colleagues identified a complication in the process. A study of consultants attempting to validate GPT-4 outputs during a problem-solving task found that the harder they pushed back against the large language model, the more it defended its answers.
The researchers said the LLM’s response, which they call “persuasion bombing,” unfolded in three escalating ways:
|
-
First, the LLM increased the intensity of its recommendation, using statistics and information to support its initial conclusion.
- After further pushback, it switched to a more overtly emotional register featuring apologies, flattering language, and renewed assurances of effort and transparency.
-
If the consultants continued to question the LLM’s results, its responses drew on a widening range of rhetorical approaches: claims about credibility, reinforcing logical arguments, and deepening rapport with the user.
|
How can you counter this? At the individual level, train employees to recognize persuasive tactics, encourage fact-checking outside the chat interface, and use prompt engineering to request neutral, academic responses. At the organizational level, consider deploying “judge agents,” LLM-based systems tasked with critiquing other AI outputs and raising counterpoints.
|
|
|
|
– Sinan Aral, Director, MIT Initiative on the Digital Economy
|
|
|
|
|
One Main Street 9th Floor | Cambridge, MA 02139 US
|
|
|
|
|